ChatGPT gets plugins and browsing

OpenAI announced plugin support for ChatGPT, giving the model, for paying subscribers on a waitlist, the ability to retrieve live information from outside its training data and to take actions through third-party services. Two plugins came from OpenAI itself: a web browser, letting ChatGPT read current pages rather than rely on its training cut-off, and a Code Interpreter that ran Python in a sandboxed environment, letting the model handle file uploads, run calculations and produce charts.

Alongside these, OpenAI opened an initial set of third-party plugins built by companies including Expedia, Instacart, Kayak, Klarna, OpenTable, Shopify, Slack, Wolfram and Zapier, plus an open-source retrieval plugin for connecting ChatGPT to a private knowledge base. In practice this meant a user could, within a single conversation, ask ChatGPT to look up a restaurant, check its opening hours on the live web, and book a table, with the model choosing which plugin to invoke and what to send it. OpenAI described plugins as “tools designed specifically for language models with safety as a core principle,” and restricted early access while it assessed misuse risks such as plugins being used to circumvent safety limits or to take unwanted actions on a user’s behalf.

The release mattered less for any single capability than for the shift in what ChatGPT was allowed to do: a system that had answered from a frozen training snapshot could now read the current web and trigger effects outside the chat window. Commentators at the time cast it as the point ChatGPT began to look less like a chatbot and more like the kind of general-purpose agent that projects such as AutoGPT, launched a week later, tried to build without waiting for OpenAI’s own plugin ecosystem to mature. Plugin adoption stayed niche, and OpenAI later folded much of the same functionality — browsing and code execution — directly into ChatGPT and, eventually, into custom GPTs.